DISQUS

Hello! Homo-Adminus Blog is using DISQUS, a powerful comment system, to manage its comments. Learn more.

Community Page

blog.kovyrin.net Jump to website »

Subscribe
- New Comments
Community
Top Commenters
- Scoundrel
  9 comments · 1 points
- dkubb
  3 comments · 1 points
- for you health
  3 comments · 1 points
- Johny
  2 comments · 2 points
- kuru
  2 comments · 1 points
Popular Threads
Recent Comments
- nice article! nice site. you're in my rss feed now ;-) keep it up
  5 months ago by for you health
  
  in Lighttpd Book from Packt - Great Thanksgiving Present
- nice article! nice site. you're in my rss feed now ;-) keep it up
  5 months ago by for you health
  
  in Dog-pile Effect and How to Avoid it with Ruby on Rails memcache-client Patch
- nice article! nice site. you're in my rss feed now ;-) keep it up
  5 months ago by for you health
  
  in ActiveMQ + Ruby Stomp Client: How to process elements one by one
- Top post. Thank you very much. :-)
  5 months ago by jamie
  
  in ActiveMQ Tips: Flow Control and Stalled Producers Problem
- Thanks for the info. May God have mercy on us all.
  5 months ago by Bulimia
  
  in Using Sphinx for Non-Fulltext Queries

Homo-Adminus Blog

Jump to original thread »

Found an Ideal I/O Scheduler for my MySQL boxes

Started by Scoundrel · 11 months ago

Today I was doing some work on one of our database servers (each of them has 4 SAS disks in RAID10 on an Adaptec controller) and it required huge multi-thread I/O-bound read load. Basically it was a set of parallel full-scan reads from a 300Gb compressed innodb table (yes, we use innodb plugin). Looking at [...] ... Continue reading »

9 comments

koticka
11 months ago

Result is predictable, this situation (raid controller) described in many documents about linux io scheduler. Scheduler (any) works only for hdd without any raid controllers.

reply

Me
11 months ago

echo 1 > /proc/sys/vm/drop_caches

reply

Yordan
11 months ago

Kernel version?

reply

Kostas Georgiou
11 months ago

Well there isn't much that the I/O schedulers can do if your access pattern is quite random beyond maybe giving you better average latencies .

Running cfq with slice_idle=0 might be a good idea btw, keeping the disks idle for up to slice_idle (8ms by default) isn't probably the best option for a db workload.

It would be interesting to get some plots/movies from seekwatcher also, I find that it gives you a nice insight on your workload and how each scheduler is affecting it.

reply

Scoundrel
11 months ago

2.6.18-53.el5

reply

Yordan
11 months ago

what is the number of threads in mysqld?

reply

Scoundrel
11 months ago

2Yordan: There were 8 separate threads in our "database load generation tool" (sphinx indexer) + some number of requests is being done from the site directly.

reply